>>>>> "Anton" == Anton Ertl <[EMAIL PROTECTED]> writes:

> David Kuehling wrote:
>>
>> one new line to the `Benchres' file:
>> 
>> 0.56 0.63 0.36 0.79 Transmeta Crusoe TM5800 (1GHz); \ gcc-2.95.4;
>> gforth-0.6.2 --enable-force-reg
>> 
>> Makes the Crusoe look as good as a P3 1000.  The compact GForth
>> engine seems to be a good target for code-morphing.

> Well, with dynamic superinstructions and replication the code is not
> that compact.  However, for these small benchmarks most of the time is
> spent in a small piece of code, which should help.  Brainless might be
> more of a challenge for the Crusoe.

When benchmarking with brainless, I just won't have any meaningful
reference data to compare to.

>> Also with dynamic superinstructions the Code-Morphing Software might
>> be able to compile indirect branches to NEXT into static branches.

> I guess you mean the indirect branches _of_ NEXT.  

That's what I actually wanted to say...

> The Crusoe would have to undo such an optimization, if writes to the
> threaded code happen, so I doubt that it does that; hmm, it could do
> that with a check of the branch target (pretty much what BTB
> prediction does).

That's what I thought as well.  The Crusoe can do some speculative
execution using a working register copy, throwing an exception if
something unexepected happens.  In theory it could even eliminate the
branch completely, putting the code-morphed primitives into consecutive
memory space.  Not sure whether it's realy that capable...
Unfortunately there are not much technical details about the Crusoe
available.

David
-- 
GnuPG public key: http://user.cs.tu-berlin.de/~dvdkhlng/dk.gpg
Fingerprint: B17A DC95 D293 657B 4205  D016 7DEF 5323 C174 7D40


---------------------------------------------------------------------
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]

Reply via email to